Automatic pipelining of elastic systems

• Autores: Marc Galceran Oms
• Directores de la Tesis: Jordi Cortadella Fortuny (dir. tes.)
• Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2011
• Idioma: inglés
• Tribunal Calificador de la Tesis: Antonio González Colás (presid.) , Josep Carmona Vargas (secret.) , Gerard Berry Berry (voc.) , Luciano Lavagno (voc.) , Alex Yakovlev (voc.)
• Texto completo no disponible (Saber más ...)
• Resumen

  • The structure of the pipeline is one of the key decisions to be made during early design stages of a sequential circuit. However, pipelining is often decided based on intuition and experience of the architects, mainly because of the significant computational costs of simulation during design space exploration and lack of analytical optimization methods capable of pipelining in presence of dependencies between iterations.

    This work presents several methods to automatically explore microarchitectures in order to find and evaluate different pipeline options. Exploration is performed by applying provably-correct by construction transformations on the microarchitecture.

    Exploration is enhanced by the capabilities of Synchronous Elastic Systems. Such systems can tolerate latency changes in computations and communications. This elasticity enables new microarchitectural trade-offs aiming at average-case optimization rather than worst case.

    The first contribution of this work is to compile a list of correct-by-construction transformations. The transformations are local, i.e., they are applied to a small part of the microarchitecture, and it is possible to pipeline a microarchitecture by applying a sequence of such transformations. Some of them can only be applied to elastic systems, since they modify the latency of the communications and computations.

    The second contribution uses elastic transformations to introduce speculative execution in an elastic system. Speculation can be used to implement some well-known architectural features like branch prediction, and can also be used to implement error detection and correction schemes.

    The third contribution uses elastic transformations to pipeline a microarchitectural description of a system. A novel framework for microarchitectural exploration is introduced, showing that the optimal pipelining of a circuit can be automatically obtained by using the previous transformations.

    Finally, a method to analyze the performance of elastic systems and obtain a lower bound of the throughput in presence of early evaluation is presented. There were already existing methods to obtain an upper bound of the throughput, and there is no known tractable method to analyze the exact throughput of an elastic system with early evaluation.